In a major advance for regenerative medicine and longevity research, scientists from OpenAI and the San Francisco-based biotech company Retro Biosciences have used a specialised artificial intelligence model to redesign key proteins that reset adult cells to a stem-cell-like state. The AI-engineered versions of the Yamanaka factors achieve more than a 50-fold increase in the expression of stem-cell reprogramming markers compared with the natural proteins, while also accelerating the process and improving DNA repair.
The findings, first detailed by OpenAI in August 2025 and subsequently validated across multiple cell types and delivery methods, continue to draw attention in scientific and biotech circles into 2026. They demonstrate how AI can rapidly explore protein designs far beyond traditional laboratory approaches.
Background: The Yamanaka Factors and Their Limitations
In 2006, Japanese scientist Shinya Yamanaka identified a set of four transcription factors—OCT4, SOX2, KLF4 and c-MYC (collectively known as OSKM)—that can reprogram mature somatic cells, such as skin fibroblasts, into induced pluripotent stem cells (iPSCs). These iPSCs resemble embryonic stem cells and can differentiate into nearly any cell type in the body. The discovery earned Yamanaka the 2012 Nobel Prize in Physiology or Medicine and opened pathways for regenerative therapies targeting conditions including blindness, diabetes, infertility and organ failure.
Despite their transformative potential, the natural Yamanaka factors are inefficient. Typically fewer than 0.1% of treated cells successfully convert, and the process often takes three weeks or longer. Efficiency declines further in cells from older or diseased donors, limiting practical applications in ageing research and clinical translation.
The AI Approach: GPT-4b Micro
OpenAI collaborated with Retro Biosciences’ Applied AI team to develop GPT-4b micro, a compact, specialised version of GPT-4o tailored for protein engineering. The model was trained primarily on protein sequences enriched with evolutionary context, biological text descriptions, co-evolutionary data and tokenised three-dimensional structural information. This training allowed it to handle both structured proteins and those with intrinsically disordered regions—features common to the Yamanaka factors, which rely on flexible interactions rather than rigid folds.
Researchers prompted the model to generate novel variants of SOX2 and KLF4, two of the four factors. The AI proposed sequences that differed from the natural human proteins by as many as 100–150 amino acids—roughly one-third of the protein chain—far more radical changes than the single-residue or few-residue mutations typical of conventional directed-evolution screens.
Laboratory Results
Wet-lab testing at Retro Biosciences confirmed the designs’ effectiveness. Over 30% of the AI-generated “RetroSOX” variants outperformed wild-type SOX2 in driving expression of early pluripotency markers such as SSEA-4. For KLF4, nearly 50% of the “RetroKLF” variants surpassed the best results from earlier manual or limited-mutation screens.
When top-performing RetroSOX and RetroKLF variants were combined with the remaining natural factors (OCT4 and MYC), the improvements became even more pronounced:
Expression of both early and late pluripotency markers (including TRA-1-60 and NANOG) rose dramatically.
Late markers appeared several days earlier than with standard OSKM.
In mesenchymal stromal cells from donors over the age of 50, more than 30% of cells expressed key markers within seven days; by day 12, colonies with iPSC-like morphology formed, and over 85% of cells activated endogenous stem-cell genes.
